Early Detection of Canola Fields Using Time-Series Sentinel-2 Imagery and Random Forest Algorithm in Google Earth Engine: A Case Study in Semi-Arid Regions of Iran

Accurate early detection of canola (Brassica napus) before the flowering stage is challenging due to significant spectral overlap with winter cereals such as wheat and barley. This study evaluates the capability of Sentinel-2 imagery combined with the Random Forest machine learning algorithm within the Google Earth Engine (GEE) platform for early canola mapping in Parsabad region, Ardabil Province, Iran.

Multi-temporal Sentinel-2 and Landsat-8 images were processed in GEE. Five spectral indices (NDVI, EVI, NDRE, CFI, and NDYI) were extracted as time series from October 2016 to May 2017. The performance of Sentinel-2 and Landsat-8 was first compared during the flowering period (April–May), where Sentinel-2 achieved superior results with overall accuracy (OA) of 0.94 and Kappa coefficient of 0.87. For early detection, classification was performed in autumn (October–December) and winter (January–March) periods. The winter period yielded the best results with OA of 0.91 and Kappa of 0.81, enabling reliable detection of canola at least one to two months before flowering. Removing March data reduced accuracy to OA 0.88, highlighting its critical role in phenological differentiation. NDRE, leveraging Sentinel-2’s Red Edge band, proved particularly effective in distinguishing canola from cereals.

This GEE-based Random Forest approach demonstrates a scalable, accurate, and cost-effective method for early crop mapping using machine learning and high-resolution satellite time series in semi-arid agricultural regions. The methodology supports timely agricultural decision-making for irrigation management and yield estimation.